Classification And Detection Of Household Waste Based On Convolutional Neural Network

Domestic waste is a misplaced resource.Proper waste separation not only increases the resource value and economic value of domestic waste,but also reduces the amount of waste and the use of treatment facilities,which has high social,economic and ecological benefits.There are many problems with the separation of domestic waste in terms of source and treatment and utilisation,such as the irregular distribution of domestic waste,the many types of waste,which are difficult to define,and the low awareness of waste separation and the low accuracy of waste separation.At the same time,the centralised classification of waste is also usually done manually,which has the problems of high staff labour intensity and low efficiency,and poses a great challenge to domestic waste classification detection.With the development of deep learning and image processing technology,target detection technology has been widely used in waste classification.In this paper,we propose two convolutional neural network-based domestic waste classification detection algorithms to effectively improve detection efficiency,in view of the problems that domestic waste is difficult to define and small targets are difficult to detect in waste images.The main work is as follows.(1)To address the problems of reduced algorithm testing effectiveness due to different sizes of household rubbish,too many types,and difficulties in rubbish definition,an SSD(Single Shot Multi Box Detector)household rubbish classification detection algorithm through feature reinforcement and upper and lower layer feature fusion is proposed.Two plug-and-play Feature Enhancement Modules(FEM)are designed to enhance the detection capability of SSD by using the idea of residual connectivity and channel selection.The experimental results show that the two feature enhancement modules and the contextual feature fusion scheme designed in this paper can improve the performance of SSD for household rubbish detection.(2)An improved algorithm based on CL-IFPN+MobileNetV2-SSD is proposed to address the problems of large number of parameters,serious memory loss and difficulty in detecting small targets in the MobileNetV2-SSD algorithm.The overall structure of the MobileNetV2 linear inverse residual bottleneck and the linkage between the convolutions are redesigned to dynamically acquire more image features;the idea of combining Non-local attention mechanism and Feature Pyramid Networks(FPN)is proposed.Implicit Feature Pyramid Network(CL-IFPN)is proposed to improve the accuracy of small target detection;the CL-IFPN is effectively fused with different feature layers of MobileNetV2-SSD to achieve real-time target classification and detection.Experiments show that the proposed CL-IFPN+MobileNetV2-SSD algorithm effectively improves the detection accuracy and detection speed for different categories of targets.To further validate the effectiveness of the proposed CL-IFPN+MobileNetV2-SSD algorithm in the field of waste classification,the algorithm was quantified and ported to the NVIDIA Xavier NX edge computing board to build a mobile intelligent waste classification bin for home use.The experiments show that the algorithm can effectively adapt to the requirements of edge computing devices on detection speed and detection accuracy,and can achieve fast and accurate domestic waste sorting detection,which has considerable engineering application value.39 figures,10 tables,79 references.